AI Lowers the Cost of Acting on Founder Conviction

Garry Tan traces two of his biggest career regrets to the same mistake: treating the market’s prevailing story as more authoritative than his own direct knowledge. That mistake matters more in an AI era because the cost of acting on idiosyncratic knowledge is falling. If a founder can build, test, and revise cheaply, the practical case for following a personal obsession rather than a consensus category gets stronger.

In 2003, after the Web 1.0 crash, Tan wanted to work at a startup but could find no Bay Area job beyond Gap IT. His choices were Microsoft’s Windows Mobile group or Expedia. He took Microsoft, at a moment when technology itself did not carry much status and the web was widely considered finished. In retrospect, he says, he was leaving precisely the area in which he had accumulated an unusual advantage—years of web-programming experience—just before social software took off.

The premise behind his “don’t LARP” formulation is not anti-strategy. It is that founders should not perform the role of the person who works on whatever appears consequential, investable, or fashionable. The question “What’s hot, and what should I work on?” substitutes market consensus for personal evidence. A better question, Tan says, is: what are you interested in, and what do you know uniquely?

Earnestness, in this account, is not naivete. It is the courage to retain confidence in direct experience when it conflicts with a blog post, a social-media consensus, or a respectable institutional view. Tan imagines the advice a time-traveling version of himself should have received in 2003: stay with the web, where you already have a five-year head start, rather than chase the next legible trend.

His refusal to join Palantir followed the same pattern. Joe Lonsdale and Stephen Cohen—his Stanford fraternity brothers—recruited him after they began working with Peter Thiel. Tan says Thiel offered him a $70,000 check, equal to his Microsoft salary, because he was convinced the new company was the right fit. Tan declined because he expected a Microsoft promotion to “level 60,” which he received. He estimates the foregone upside now as a $2 billion to $4 billion mistake.

What he missed was available in the territory, not on the map. Lonsdale and Cohen were among the smartest people he knew; their work addressed a visible gap between what strong computer scientists in Palo Alto could build and what government agencies had access to. But Tan was reasoning from what he believed a good investor would find cool, rather than from the people and facts directly in front of him.

That is why Silicon Valley’s important ideas often begin at the fringe, he argues. The proposition that everyone should have a computer once belonged to “fringe, weirdo punks,” not the center of institutional power. Acharya adds a test for whether an area is genuinely interesting: one may find the edge of the public conversation, but not its intellectual edge. The more one investigates, the more remains to understand.

The internet has made it easier for people with strange, deep interests to find one another quickly—through communities on Reddit, X, and elsewhere—rather than waiting for a physical scene to validate them. The relevant social question is not whether a pursuit looks mainstream. It is whether the people involved are working from an unusual truth and can build around it.

YC’s role, in Tan’s telling, is to make that search for one’s people less dependent on social access. He credits Paul Graham, Jessica Livingston, Trevor Blackwell, Robert Morris, and the other original YC figures with recognizing that intelligence and building ability were more widely distributed than access to Silicon Valley’s network. The response was procedural: an application that began as 12 questions and later added a one-minute video, evaluated by builders trying to give applicants a fair hearing rather than by what YC calls “scensters.”

Tan says YC now has 16 equal partners, all of whom went through the program and had successful outcomes. Its recent Startup School brought 7,000 people to San Francisco, most visiting for the first time. He calls the proposition “birthright for tech”: founders should be assessed on their ideas and ability to execute, not on their proximity to the right schools, parties, or networks.

Access, in Tan’s view, also needs to include candor. Founders need people with whom they can speak honestly on the day they lose their best customer, their best engineer quits, or a co-founder loses hope. Silicon Valley’s useful community is not simply a distribution channel for status. It is a place where someone can say what is failing without having to maintain the appearance of “killing it.”

? anish-acharya argues that code is no longer precious in the way it once was. A product idea previously implied a product manager, a specification, engineers, QA, and enough organizational conviction to justify the coordination burden. That friction favored projects that looked important from the outset, even though consequential ideas often begin as apparently trivial interests.

Tan’s response is that founders should use the new production capacity to follow those interests, build, and learn rather than wait for permission or certainty. His own g-stack and g-brain experiments began as exploration. Seeing friends return to coding made him think something had changed, so he began making things.

Recording YC office hours gave him several months of material from which he says he identified rhetorical methods partners use to induce a step change in founders’ ambition. He put those methods into a markdown file—a script for advising founders—and asked Claude to reduce the intensity of the prompt by 90% before open-sourcing a version. The more forceful version, he jokes, remains part of the value of going through YC.

The important habit is not merely consuming AI products. It is developing agency and taste by repeatedly using them. Tan does not think those qualities are fixed traits held by a naturally high-agency minority. Some people may start with more of them, he says, but they can be practiced. His own record of leaving web programming at the wrong time and declining Palantir is, for him, evidence that people can inspect their mistakes, reason through them, and make different choices.

AI lowers the cost of that practice. A person can have an idea, make a version, generate tests, invoke QA, inspect the result, and revise. Much of the work can be low-stakes and private. The immediate output need not become a company; making trivial things can create the intuition needed to recognize a serious opportunity later.

That expanded individual capacity changes, but does not erase, the case for co-founders. Garry Tan still thinks co-founders are net positive when all else is equal. A second person joining an unusual conviction makes it a shared undertaking rather than a solitary fixation. But with agentic coding and “vibe coding,” he says, a founder can be “400 of that person.” The implication is not that founders should merely build the companies that were accessible to earlier generations with smaller teams. They should be more ambitious.

It also means that pure SaaS may no longer be a sufficient destination. Tan says a pure pre-seed SaaS company may not clearly exist in the same form five or 10 years from now. SaaS can still be a wedge, but, in his forecast, a founder starting one in 2026 needs to know what comes after it: a moat in data, network effects, or another advantage that cannot simply be recreated by generating more code.

Two years earlier, he says, a 10-to-20-times next-twelve-months-revenue multiple for SaaS could appear to be an “iron law.” It no longer does. The difficulty of building software cannot itself supply a durable advantage; Tan’s concern is whether a company compounds what it learns about users, operations, and its domain.

Garry Tan sees the central opportunity in agents not as faster coding alone, but as the conversion of recurring work into reusable operational judgment.

His term for the process is “skillification.” A founder completes a difficult task, then turns the result into a markdown file, code, and tests that can be reused and scheduled. The first version is likely to be bad and expensive. But when the founder can identify errors quickly and plainly—this was wrong; fix that—the agent’s trace and history can be incorporated into a better skill file. Future failures become bug fixes rather than fresh work.

The analogy is intentionally broad. Tan says the model applies to sales, marketing, customer support, engineering, and any recurring business process. A skill file can run a job consistently, at scale, and without requiring the organization to rediscover the method every time.

His own engineering work became an example. In building g-stack, Tan says, he automated product-management thinking, engineering-management checks for unit and end-to-end test coverage, and browser-level QA. Eventually, after automating most other steps, he found himself spending his own time on black-box testing. That revealed the next bottleneck. Building agentic systems, he says, is largely the practice of locating the current constraint and creating software or an instruction set that removes it.

Tan describes that as a different unit of company-building: not a collection of employees performing disconnected tasks, but an accumulating library of skills, tests, data, and procedures that can be run again. He says YC is seeing companies reach zero to $15 million in annual recurring revenue in about four months with two or three people, hundreds of agents, and a few hundred skill files.

The premise is not that an agentic company runs indefinitely without supervision. As the number of files and the volume of data grows, Tan says, provenance and conflict management become essential. If two facts contradict one another, the system needs a basis for deciding which one wins—such as recency or source authority—and scheduled processes to sweep for errors.

This is where Tan sees an advantage for founders in their 30s and 40s. The person who has built substantial engineering systems knows where information goes stale, where exceptions accumulate, and what requires monitoring. AI can multiply that accumulated judgment. He expects younger founders of exceptional ability to remain, but argues that a person “who’s been around the block” can now operate at a scale that previously required an entire department.

? anish-acharya pushes the argument beyond workflow automation. The more consequential business loops, he suggests, are loops in which an organization’s output changes how the business itself operates. In that setting, agents can make disagreements and bottlenecks easier to inspect without carrying the human emotion that often makes conflict costly.

Tan’s example is Pedro of Brex. Pedro, Tan says, introduced him to OpenDevin and built an open-source layer called CrabTrap that watches the agent’s network traffic and other activity. According to Tan, this made it possible to use the system in a heavily regulated fintech setting despite concerns that OpenDevin could be unsafe.

The notable application is not only engineering work. Tan says Pedro uses agents to examine meeting transcripts from direct reports, including material two levels down. The system can identify broken processes, disputes, and unresolved conflicts in meetings Pedro did not attend. It can give him detailed context from several weeks of a particular team’s work before he enters a meeting, enabling a quick decision when he believes one side’s argument is stronger.

For Tan, this addresses a defining management problem: a business eventually becomes too large to fit inside one person’s head. Conventional organizations cope with that limit through reporting layers, handoffs, and managerial systems that lose context along the way.

An agent with memory and retrieval can hold far more relevant context, Tan says—the equivalent of “three Harry Potter books”—and retrieve a precise slice when needed. His claim is not that the system resolves organizational disputes reliably on its own. It is that it can make the underlying record more legible to the person who still has to exercise judgment.

Tan’s memory of Windows Mobile illustrates the existing failure mode. His team needed integration from the Windows group across the highway in Redmond. Emails went unanswered; bugs were not fixed or even marked “won’t fix.” After spending hours trying to address a blocker affecting roughly a thousand users, Tan and a product-management mentor went to the other team carrying a baseball bat. He says it was not intended for violence, but a bizarre gesture of pressure.

The underlying problem was not necessarily malice. The person who could resolve the issue may simply have lacked the context to know how much time another team was losing. Seven layers of hierarchy can separate a problem from the person able to fix it.

A system that maps who is blocked, what dependencies exist, and where work is failing could surface those problems directly. Tan imagines work that formerly took six months being released in a day or two. Executives still set direction, and people doing the work retain local knowledge. But the middle layers devoted to routing information, organizing tasks, and handling routine coordination need not remain exclusively human.

That is why Tan says AI may erase what Venkatesh Rao calls the boundary between being above and below an API line. He points to Clawvisor, a YC-funded API company, which created a bug-report endpoint meant for agents. When an agent encounters a defect, the endpoint can respond in real time that the report has been filed, offer a likely fix timeline, and suggest a workaround. Rather than a one-way interface that tells a worker what to do, the system can participate in an exchange grounded in the work being done.

Tan connects this to the Toyota production system as he understands it: the worker on the line should be able to change the line because that worker has the most context about how to improve the task. Agents can make more organizational context available, but the point is to make human work more responsive, not simply more controlled.

A startup, Tan says, can organize itself around such systems in ways a large incumbent may not be able to. In his view, every startup must.

Garry Tan does not take rapid technical progress to imply rapid institutional replacement. His “white pill” is that bureaucracy, limited attention, coordination problems, and middle management will slow change considerably.

This is partly a rebuttal to the view that AI will quickly eliminate white-collar work and create a permanent underclass. Companies, governments, and social institutions are built around limited human beings, he argues. Their inertia is real; large incumbents have structural moats; and a startup will not simply replace every major company. The transition, in his estimate, is closer to 20 years.

That timeline also reflects generational change. Tan looks at the 18-to-22-year-olds who attended YC’s Startup School and sees people who became adults in the era of ChatGPT. Their expectations of software and organizations will eventually shape institutions, much as web- and mobile-native people eventually moved into positions of authority. But the intervening process is slow.

The same tension shapes his forecast for consumer AI. The computer’s immediate form factor may persist for a while, but it is unlikely to remain stable. Voice will matter, as will durable memory and systems that ingest relevant information, distinguish speakers, use computers, and maintain an ongoing model of the user.

The desired assistant is not just a chat interface. Tan imagines something that knows a person’s hopes, fears, and desires and persistently looks for ways to help. Current users may be satisfied with ChatGPT, Claude, or similar systems, but expectations will rise as people want more context and continuity than a standard chat window provides.

He expects competition over that layer—the “harness wars”—to emerge around 2027. The immediate constraint is cost: full-strength agentic use can currently cost a founder $50,000 to $100,000 a year, which is incompatible with broadly giving consumers expensive inference as a free trial.

Acharya makes the consumer-cost argument more directly. Software’s historic magic, he says, has been near-zero marginal distribution; AI products cannot replicate that model while inference remains expensive. Yet the price of a frontier model can remain high even as the model that was frontier shortly before becomes cheap. He sees that cost progression as an opening for more ambitious consumer AI, including products with emotional texture rather than another generation of static software.

Tan agrees that the opportunity is early, but has revised his own pace of expectation. Open agent systems initially convinced him that the transition would happen immediately. They no longer do. Powerful capabilities may already change how small companies are built, while consumer habits and institutional redesign arrive on a longer timetable.

Tan applies a similar preference for direct engagement to local politics, though he treats the obstacles there as human-coordination problems rather than intelligence-bound ones. Garry Tan says his civic “white pill” is believing that rule of law still operates and that people in government are trying, however imperfectly, to help. Jury duty in San Francisco made that belief concrete for him: he saw an orderly process involving a cross-section of the city rather than an abstraction about institutional failure.

His engagement became personal during COVID. Tan says Asian Americans make up roughly 25% of San Francisco’s voting population and around 30% of its population. He was angered by what he describes as a district attorney turning a blind eye to serious crimes against Asian elders and a school board hostile to Asian-American students who wanted to take algebra in public middle school.

The algebra issue connected directly to his own path. Tan attended public school in Fremont and believes that without middle-school algebra he could not have taken calculus as a senior, studied engineering at Stanford, or reached his subsequent career. He came to see those fights through a builder’s view of incentives, supply and demand, housing, and infrastructure.

His proposed response is organization rather than resignation: vote, pay attention, hold media as well as government accountable, and accept personal or reputational costs when institutions fail. Tan says Chinese residents on San Francisco’s west side organized around access to algebra and violence against elders, while local television news failed to cover anti-Chinese and anti-Asian crime adequately.

He is less interested in being consumed by national politics than in housing, crime, treatment, recovery, and the conditions affecting people nearby. Fixing local institutions, he argues, can produce better state and national leadership over five- and 10-year horizons because people who rise through functional systems carry those lessons with them.